ConceptGate: Few-Shot Concept Reading and Steering in Frozen Language Models

Language models encode many concepts as directions in their activations, so one direction can both detect a concept and steer the model. Operators of deployed models want both, often from a few examples and without retraining. We test how far this goes with one residual-stream direction per concept, fit in closed form from about ten examples per class. On five models up to 2.6B parameters, we compare it with a logistic probe on the same layers, matched-norm random directions, and random selection. We report three findings. First, steering along the few-shot direction moves refusal both up and down. On Qwen2.5-0.5B, refusal is 64% unsteered, 35% when we steer away, and 79% when we steer toward. Second, the direction reads the surface form of an attack, not its intent. On two models and on 1,725 real prompts, a gate on the direction fires on 93% of harmless commands to the model and on 53% of attacks written as ordinary requests. Its false-positive rate changes from 18% to 96% with the register of benign traffic. Third, steerability is predictable per prompt. On Qwen2.5-0.5B, the unsteered activations predict how far a fixed write moves a prompt's sampled refusal rate (Spearman rho = 0.58, permutation z = 6.7). As a steering trigger, this predictor beats random selection at every tested coverage. The concept gate reads the same quantity with the wrong sign (rho = -0.30). On gemma-2-2b the write barely moves refusal, so this claim is untested there. The controls also show where the design adds nothing: the detector matches the logistic probe, and depth fusion and mixture densities give no gain on real models. Two lessons for steering evaluations follow. A missing chat template reversed the sign of an early result, and an unvalidated outcome measure gave a false null. Code: https://github.com/NISH1001/conceptgate.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24287378
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

ConceptGate: Few-Shot Concept Reading and Steering in Frozen Language Models

Nishan Pantha
2 links
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
preprint

ConceptGate: Few-Shot Concept Reading and Steering in Frozen Language Models

Nishan Pantha
preprint en

Abstract

Language models encode many concepts as directions in their activations, so one direction can both detect a concept and steer the model. Operators of deployed models want both, often from a few examples and without retraining. We test how far this goes with one residual-stream direction per concept, fit in closed form from about ten examples per class. On five models up to 2.6B parameters, we compare it with a logistic probe on the same layers, matched-norm random directions, and random selection. We report three findings. First, steering along the few-shot direction moves refusal both up and down. On Qwen2.5-0.5B, refusal is 64% unsteered, 35% when we steer away, and 79% when we steer toward. Second, the direction reads the surface form of an attack, not its intent. On two models and on 1,725 real prompts, a gate on the direction fires on 93% of harmless commands to the model and on 53% of attacks written as ordinary requests. Its false-positive rate changes from 18% to 96% with the register of benign traffic. Third, steerability is predictable per prompt. On Qwen2.5-0.5B, the unsteered activations predict how far a fixed write moves a prompt's sampled refusal rate (Spearman rho = 0.58, permutation z = 6.7). As a steering trigger, this predictor beats random selection at every tested coverage. The concept gate reads the same quantity with the wrong sign (rho = -0.30). On gemma-2-2b the write barely moves refusal, so this claim is untested there. The controls also show where the design adds nothing: the detector matches the logistic probe, and depth fusion and mixture densities give no gain on real models. Two lessons for steering evaluations follow. A missing chat template reversed the sign of an early result, and an unvalidated outcome measure gave a false null. Code: https://github.com/NISH1001/conceptgate.

Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon